In generative modeling, standard measure theory adopts the a.e. convention, which legitimately ignores any error component carried by a null-support set: a single null element can hide out-of-domain error inside null support and never be explicitly accounted for. The contribution of this paper is not a new distributional distance, but rather a verifiable structure that writes down which underlying predicate supports a high-level object, when it is invertible, and at which layer the error resides, from which we propose the enumerating posterior-supported prior framework. The framework operates on a three-layer categorical chain Quot->MorTop->Set: the first two layers posit no point elements and take binary-predicate quotient partitions as objects; the replica-index counting measure successively discriminates one-to-many collections into injections, lifting object collection from surjection to a quotient-layer bijection; the mutual exclusion of preimages in the shared measure space is resolved by an invertible MPT. Four results consolidate into Thm. 7: binary quotient-partition meta-categorical objects support the set categorical functional along categorical support via measure descent--recovery, enumerating the morphism itself rather than its preimage and object; their multiplicity is normalized through replica-index injection and accounted as an explicitly first-order error. The contribution of this paper is a transformation from nontrivial-input construction -> engineering-computable first-order output: the binary quotient-partition meta-categorical objects, supported via categorical support measure descent--recovery, constitute a substantive nontrivial input---precisely because they lift the out-of-domain error that hid inside null support into a first-order falsifiable coefficient mu_err, a finitely obtainable finite integral first-order error.
Figures & tables
Route
Invertible / pairing
Error tracking
Open problem solved here
Normalizing flow flow matching
point-wise invertible, rep.-dependent
Jacobian, no replica tracking
null-embedding ambiguity
Optimal transport metric comparison
coupling, not predicate bijection
only volume diff.
training–generation pairing bias
Diffusion model
not aimed at invertible pairing
no first-order error layering
iterative high-order error accumulation
Ours
quotient-layer bijection + invertible MPT
explicit first-order accounting, 3 channels close
all three
Table 1: Route comparison: invertibility, error tracking, open problems.
Layer
Object
Point elem.
Measure source
Failure if removed
Quot
binary predicate / quotient partition
none
none
Null-embedding ambiguity recurs
MorTop
quotient block / differential quotient
none
induced by partition or bijection
First-order error degenerates to high order
Set
derived instance
yes
pre-given
No accountable reading
Table 2: Three layers of the categorical chain.
Figure 1: Forward-propagated zero-value saddle point and backward-propagated random saddle point align at the zero-value saddle point.
finite preimage mutual-exclusion constraint (see Corollary 10a)
no point elements; Boolean decision only
MorTop mapping category
infinite predicate
layered mutually-exclusive additive group Gexcl
quotient-class bijection; measure via descent
Set set category
numerical functional
measure-space topological mutual exclusion
derived instances; dual measure
Table 3: Three layers of the categorical chain and their mutual-exclusion forms.
Figure 2: Gaussian functional and Gaussian behavior (fiber bundle): (a) Gaussian marginal distribution (colored by partition segment); (b) shuffled Gaussian—a permutation variant of the same marginal segment, topologically equivalent to (a) under the marginal-distribution reading.
Figure 3: Truncated Gaussian and reparameterized Gaussianization. Top: sampling truncated at the zero-value saddle point (one-sided integral domain); bottom: disentangled after Gaussianization; bottom band: aligned-but-non-uniform vs. non-aligned-but-uniform configurations, measured by the block-share/Lebesgue-share difference ( ΔI , D).
Conventions 1–4 architectural; differential reading needs no Assumption 1
④
Principle 4 + Theorem 6(i) + Condition 1 → Lemma 6a(3) existence of a gradient direction decreasing first-order error (monotonic non-increase not guaranteed)
Lemma 6a, B.109
Condition 1 is an implementation-layer per-round condition (registered in B.4), not a mathematical assumption; monotonicity claim abandoned
midpoint selection is definitional selection (see Definition 25c; three points: boundary midpoint / subtracted saddle =x0sad , integral boundary =S0Hil , gradient direction =Smid )
Table 8: Premise-dependency table for cross-layer invocation chains (only direct premises are recorded).
Route
Main reading
Invertibility/ pairing
Structural difference
Our improvement
Normalizing flow
Invertible transport and Jacobian
Point-wise invertible; sample pairing still depends on representation
Does not handle surjection-cover completeness of the element-free quotient layer
Null-embedding ambiguity : surjective morphisms, after being successively discriminated into injections by the replica-index measure, are lifted to a quotient-layer bijection; predicate matching gives a bijection criterion [ 37 ; 8 ] [*1]
Optimal transport / metric measure comparison
Displacement, kinetic energy, or structural distance
Usually yields couplings rather than predicate bijections
Does not track replicas via quotient-block predicate words
Iterative high-order error accumulation : first-order error is explicitly localized; high-order channels close via three-channel verification (see Proposition 23c) [ 3 ; 33 ; 39 ; 40 ; 16 ] [*3]
Ours
Quotient-partition predicates and dual-measure descent–recovery
Quotient-layer bijection and invertible MPT
Uses categorical support instead of point-wise representation to organize the three-layer relations
The three problems obtain a structural resolution at the same reading layer, and the sole backbone mathematical assumption is Assumption 1 (implementation-layer conditions are registered separately and do not enter the backbone assumption list)
Table 9: Item-by-item comparison with existing routes (see §1 related work).
Item
Standard result (citation/basis)
Our new treatment / new proposition
Theorem 1 (quotient equivalence class ⇔ element-disjoint ⇔ no collapse ⇔Set functional well-defined)
Main statement: pairwise disjoint quotient-block family (no collapse) ⇔ equivalence-class differential definable as a Set functional; the finite equal-cardinality special case reduces to “no collapse ⇔ quotient-set bijection ⇔ block-pairing bijection”
Table 12: MNIST clustering performance and latent dimension.
Figure 4: MNIST training indicators: left is 60000 training / 10000 test, right is 10000 training / 60000 test; the blue line is the Lebesgue excess ΔLeb (the Lebesgue area difference between the mid-layer zero-value saddle point and the outer variational Gaussian vertex); display magnification see each panel legend.
In applied statistics and machine learning, the gold standards used for training are often biased and almost always noisy. Dawid and Skene's justifiably popular crowdsourcing model adjusts for rater sensitivity and specificity, but fails to capture distributional properties of rating data gathered for training, which in turn biases training. In this study, we introduce a general purpose measurement-error model with which we can infer consensus categories by adding item-level effects for difficulty, discriminativeness, and guessability. We further show how to constrain the bimodal posterior of these models to avoid adversarial raters. We validate our model's goodness of fit with posterior predictive checks, the Bayesian analogue of χ2 tests, and assess its predictive accuracy using leave-one-out cross-validation. We illustrate our new model with two well-studied data sets, binary rating data for caries in dental X-rays and implication in natural language.
Seong Woo Han, Ozan Adıgüzel, Bob Carpenter
University of Pennsylvania · Columbia University · Center for Computational Mathematics, Flatiron Institute
Large language models (LLMs) are increasingly used as conditional generators for structured data, relying on in-context learning (ICL) to adapt to new distributions without parameter updates. We investigate the limits of ICL for structured generation under distribution mismatch, using high-cardinality tabular data as a controlled test case, and identify a structural failure mode we term categorical prior lock-in: the inability of ICL to update the model's prior over token distributions inherited from pre-training. Across two 7B-parameter open-weight models, ICL improves numerical fidelity with additional examples but exhibits a sharp ceiling on categorical distributions, failing to reproduce rare classes entirely. Parameter-efficient fine-tuning (LoRA) overcomes these limitations but introduces measurable memorization risk and, in some cases, destabilizes structured output generation, highlighting a fundamental trade-off between adaptability and privacy.
Antonio Pelusi, Stefano Braghin, Alberto Trombetta
University of Insubria, Italy · IBM Research, Ireland
Learned priors are used to reconstruct regions that measurements barely reach, and to report uncertainty there. They can fail silently: the reported interval excludes the truth while every ground-truth-free diagnostic stays within its noise. We call such errors the blind set and ask whether they can be predicted before the target data are seen. For a correctly specified linear-Gaussian model averaged over the prior, the risk of a silent failure factorises exactly into non-detection and non-coverage, which does not hold for learned systems in general; we build a score that approximates this reference by linearising a VAE prior. In a prospective study on real X-ray CT objects (2DeteCT) under simulated restricted access, with predictions committed by hash before any trial, the score ranked the realised blindness of two learned inference systems (latent Laplace and ensemble Kalman inversion) with Spearman ρ=0.79 and 0.84 over 240 pre-specified errors. Simple structural predictors would also have passed these tests. Sequential post-hoc analyses, each specified before its own computation, show that under bounded reach the learned score gave no detectable improvement over a physics-based baseline, in rankings or trial-level log scores. At the weakest attenuation studied, prior information improved rankings over the physics-based baseline, most clearly for ensemble Kalman inversion, but a Gaussian prior did as well and the gain was not detectable at the trial level. The score is not a calibrated probability. Silent failures can be anticipated here, but their structure follows mainly from measurement access and the decision quantity rather than from the learned prior.
Mario Gianni
School of Computer Science and Informatics, University of Liverpool, UK